Radiation prediction correction method and device based on multi-source data fusion and feature distillation

Abstract

Embodiments of the present application relate to the technical field of meteorological forecasting, new energy power prediction and artificial intelligence cross, and particularly relate to a radiation forecasting correction method and device based on multi-source data fusion and feature distillation. A specific implementation of the method includes: obtaining a reanalysis data set, a site observation data set and a historical NWP data set; based on the site observation data set, performing physical correction processing on each reanalysis data in the reanalysis data set to generate privileged information, thereby obtaining a privileged information set; based on the historical NWP data set and the privileged information set, training a multi-source fusion correction model; pre-processing real-time NWP data obtained to generate real-time NWP pre-processed data; inputting the real-time NWP pre-processed data into the pre-trained multi-source fusion correction model to obtain a high-precision radiation correction forecast field. The implementation can obtain a high-precision, high-spatial-field-quality and well-preserved-detail correction subdivided resolution field.

Description

Technical Field

[0001] The embodiments of this application relate to the interdisciplinary fields of meteorological forecasting technology, new energy power prediction and artificial intelligence, specifically to a radiation forecast correction method and apparatus based on multi-source data fusion and characteristic distillation. Background Technology

[0002] Solar radiation is a core element of the Earth system's energy balance, and accurate forecasting of surface solar radiation is crucial for the development of renewable energy. With the continuous development of numerical weather prediction models, radiation forecasting based on physical processes has become a primary means of operational forecasting. However, numerical weather prediction models inherently suffer from parameterization errors when handling cloud microphysical processes and aerosol radiative transport, leading to significant systematic biases in their output radiation forecasts. For example, at the edges of convective cloud clusters or in areas with complex topography, numerical forecasts often exhibit "missed" or "false" predictions, resulting in persistently high root mean square errors.

[0003] To correct these biases, meteorologists have developed various post-processing techniques. Early studies primarily employed statistical correction methods based on station observations. These methods established statistical relationships between model outputs and station observations, then used spatial interpolation to extend the correction results to a gridded field. While this approach performs well in densely populated areas, in sparsely observed regions, especially those with complex terrain, the strong spatial heterogeneity of meteorological elements means that simple spatial interpolation can lead to significant error propagation. Furthermore, traditional bias correction methods often neglect the spatial heterogeneity of radiation biases, resulting in overly smooth spatial textures in the corrected field, losing crucial cloud movement characteristics essential for short-term photovoltaic forecasts.

[0004] As reanalysis data becomes increasingly sophisticated, correction methods utilizing reanalysis data as a reference benchmark are gaining popularity. New-generation reanalysis datasets such as ERA5 (the fifth-generation global atmospheric reanalysis dataset) integrate multi-source observations through data assimilation techniques, providing globally covered datasets with uniform spatiotemporal resolution (e.g., ERA5 is 0.25° × 0.25°, output hourly). This facilitates cross-regional and long-term series analysis and is widely recognized as the gridded dataset closest to the "true state of the atmosphere." Therefore, many studies directly treat these reanalysis data as "true values" to build correction models. However, the assimilation process of reanalysis data relies on the physical framework of the numerical model. If the model does not adequately represent the physical processes, it can lead to systematic biases in the reanalysis data, which can be further amplified in subsequent forecast corrections. Furthermore, the generation of ERA5 depends on a complex four-dimensional variational assimilation system, requiring the collection and quality control of global observational data. This results in a typical 3- to 5-day time lag in the release of ERA5 and other similar datasets.

[0005] In recent years, deep learning-based image-to-image conversion methods have provided new approaches for addressing the correction problem of spatially continuous grid fields. However, existing research faces two major bottlenecks: first, training relies on multi-source data (such as forecast fields and reanalysis fields), while real-time acquisition of observation data and ERA5 is often impossible during operational deployment, leading to model failure or even failure to run; second, how to effectively utilize sparse ground truth observations from various sites to supervise the training of the entire grid field while maintaining physical plausibility. Summary of the Invention

[0006] The summary section of this application is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description section below. This summary section is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0007] Some embodiments of this application propose a radiation forecast correction method, apparatus, computer device, and computer-readable storage medium based on multi-source data fusion and feature distillation. The core idea is to use high-quality past data (privileged information) to teach the model during training, and to reproduce high-quality physical features using only real-time data during inference. This addresses one or more of the technical problems mentioned in the background section above.

[0008] In a first aspect, some embodiments of this application provide a radiation forecast correction method based on multi-source data fusion and feature distillation. The method includes: acquiring a reanalysis dataset, a site observation dataset, and a historical NWP dataset; performing physical correction processing on each reanalysis data in the reanalysis dataset based on the site observation dataset to generate privileged information, thereby obtaining a privileged information set; training a multi-source fusion correction model based on the historical NWP dataset and the privileged information set; preprocessing the acquired real-time NWP data to generate real-time NWP preprocessed data; and inputting the real-time NWP preprocessed data into the pre-trained multi-source fusion correction model to obtain a high-precision radiation correction forecast field.

[0009] Secondly, some embodiments of this application provide a radiation forecast correction apparatus based on multi-source data fusion and feature distillation. The apparatus includes: an acquisition unit configured to acquire a reanalysis dataset, a site observation dataset, and a historical NWP dataset; a physical correction unit configured to perform physical correction processing on each reanalysis data in the reanalysis dataset based on the site observation dataset to generate privileged information, thereby obtaining a privileged information set; a training unit configured to train a multi-source fusion correction model based on the historical NWP dataset and the privileged information set; a preprocessing unit configured to preprocess the acquired real-time NWP data to generate real-time NWP preprocessed data; and an input unit configured to input the real-time NWP preprocessed data into the pre-trained multi-source fusion correction model to obtain a high-precision radiation correction forecast field.

[0010] Thirdly, this application also provides a computer device, which includes a processor, a memory, and a computer program stored in the memory and executable by the processor, wherein when the computer program is executed by the processor, it implements the method described in any implementation of the first aspect above.

[0011] Fourthly, this application also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements the method described in any implementation of the first aspect above.

[0012] The various embodiments of this application have the following beneficial effects: Firstly, the radiation forecast correction method based on multi-source data fusion and feature distillation, as described in some embodiments of this application, effectively fuses multi-source information and learns under the supervision of station ground truth. This reduces the mean absolute error (MAE) and root mean square error (RMSE) of the corrected radiation forecast at the validation stations, while increasing the correlation coefficient, thus significantly improving forecast accuracy. Secondly, the innovative feature distillation mechanism ensures the reliability of the model in real-world operational scenarios, overcoming the bottleneck of inconsistent training and deployment data, and enabling the automatic generation of high-precision correction products for operational use. Furthermore, the multi-source fusion correction model outputs a spatially continuous and physically reasonable gridded field, overcoming the shortcomings of traditional point-to-point correction methods in field structure, and exhibiting reasonable performance even in areas without stations. Next, the gradient loss and other design features enable the model to retain valuable small-scale weather details from the original high-resolution forecast while correcting biases. Therefore, a high-precision, high-quality spatial field with good detail preservation can be obtained for the corrected fine-resolution field.

[0013] The technical solution of this application includes the following key innovations:

[0014] 1. Introducing Physically Consistent Residual Kriging Preprocessing: To ensure the model learns the most accurate physical laws, this application does not directly use the original ERA5 data as labels, but instead utilizes data from high-precision ground-based radiation observation stations for correction. Unlike conventional numerical correction, this application employs a residual kriging algorithm. This algorithm first calculates the residuals between the observed values ​​and the ERA5 background field, and then fits the spatial autocorrelation of the residuals using a variogram function. This method effectively diffuses the error information from discrete stations to the entire field in a manner consistent with geostatistical laws, generating a reanalysis correction field with extremely high physical consistency, serving as privileged information during the training phase.

[0015] 2. Privileged Information Distillation Architecture Based on Modal Dropout: This application constructs a teacher-student network architecture: the teacher network's input consists of NWP forecasts and reanalysis correction fields. Because the input includes a correction field containing observational information, the teacher network can capture atmospheric conditions with extremely high accuracy. The student network's input is only the NWP forecast. During training, feature distillation loss is used to force the intermediate layer feature maps of the student network to approximate the feature maps of the teacher network. Simultaneously, a modal dropout mechanism is introduced, randomly cutting off the privileged information input of the teacher network. This forces the model to reconstruct deep physical features (such as cloud thickness distribution and water vapor transport channels) that would otherwise require reanalysis data, even with only NWP input, through nonlinear mapping within the network. This mechanism ensures that the model maintains near-dual-source fusion high-precision performance during the inference phase when real-time reanalysis data is unavailable.

[0016] 3. A composite loss function incorporating physical awareness was constructed. Addressing the high-frequency abrupt changes in the solar radiation field at cloud edges—mathematically equivalent to "edges" in an image, but geographically representing discontinuities in radiative flux—traditional mean square error (MSE) tends to generate a blurred mean field, smoothing out these discontinuities and distorting the physical process. This invention innovatively introduces spatial gradient difference loss into the loss function. By penalizing the inconsistency in gradient between the predicted field and the original NWP field, it forces the model to retain the spatial texture details brought by the weather system. This is not merely a technology transfer but also a way to satisfy physical conservation laws. Furthermore, sparse site loss ensures the model has absolute numerical accuracy at geographically recorded locations. Attached Figure Description

[0017] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0018] Figure 1This is a flowchart of some embodiments of the radiation forecast correction method based on multi-source data fusion and characteristic distillation according to this application;

[0019] Figure 2 This is a technical flowchart of some embodiments of the radiation forecast correction method based on multi-source data fusion and feature distillation according to this application;

[0020] Figure 3 This is a schematic diagram of the residual kriging correction principle according to some embodiments of the radiation forecast correction method based on multi-source data fusion and characteristic distillation according to this application;

[0021] Figure 4 This is a model structure diagram of a multi-source fusion correction model based on some embodiments of the radiation forecast correction method based on multi-source data fusion and feature distillation according to this application;

[0022] Figure 5 This is a comparison diagram of the data flow during the training and inference stages of a multi-source fusion correction model according to some embodiments of the radiation forecast correction method based on multi-source data fusion and feature distillation in this application;

[0023] Figure 6 It is a comparison chart of the original forecast, traditional MOS, ordinary CNN, radiation forecast correction method based on multi-source data fusion and feature distillation, and the evaluation results assuming that ERA5 can acquire them in real time;

[0024] Figure 7 These are schematic diagrams of some embodiments of the radiation forecast correction device based on multi-source data fusion and feature distillation according to this application.

[0025] Figure 8 This is a schematic diagram of the structure of a computer device suitable for implementing some embodiments of this application. Detailed Implementation

[0026] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While some embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this application. It should be understood that the drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0027] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0028] It should be noted that the concepts of "first" and "second" mentioned in this application are only used to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0029] It should be noted that the terms "a" and "a plurality of" used in this application are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0030] The names of the messages or information exchanged between multiple devices in the embodiments of this application are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0031] The present application will now be described in detail with reference to the accompanying drawings and embodiments.

[0032] Figure 1 A flowchart 100 is shown illustrating some embodiments of the radiation forecast correction method based on multi-source data fusion and characteristic distillation according to this application. This radiation forecast correction method based on multi-source data fusion and characteristic distillation includes the following steps:

[0033] Step 101: Obtain the reanalysis dataset, site observation dataset, and historical NWP dataset.

[0034] In some embodiments, the entity executing the radiation forecast correction method based on multi-source data fusion and feature distillation can acquire reanalysis data, station observation data, and historical NWP data at each time granularity within a preset time period from a storage terminal via wired or wireless connection, resulting in a reanalysis data set, a station observation dataset, and a historical NWP (numerical weather prediction data) dataset. The reanalysis data in the reanalysis data set may include, but is not limited to, at least one of the following: radiation, temperature, air pressure, specific humidity, total cloud cover, and low cloud cover. The reanalysis data in the reanalysis data set can be a spatiotemporally consistent gridded historical meteorological dataset generated by combining multi-source observation data with numerical models through data assimilation techniques. For example, the reanalysis data in the reanalysis data set may be an ERA5 (fifth-generation global atmospheric reanalysis dataset) reanalysis product (resolution 0.25°×0.25°, 1-hour interval). The station observation data in the station observation dataset may be hourly surface solar radiation observation data from approximately 300 national radiation stations. The station observation data in the station observation dataset may include: radiation. The NWP data in the historical NWP dataset can be surface solar radiation shortwave flux forecast products (spatial resolution 0.1°×0.1°, temporal resolution 1 hour) using the High Resolution Forecasting Model (HRES) for a preset time period. The NWP data in the historical NWP dataset can include: radiation, temperature, air pressure, specific humidity, total cloud cover, and low cloud cover. Here, the reanalysis data in the reanalysis dataset and the NWP data in the historical NWP dataset represent gridded data. The station observation data in the station observation dataset represents station data. A grid can represent: a discretized spatial unit used to represent a continuous radiation field. A station can represent: sparse locations of actual observations. Here, the station observation data in the station observation dataset can include observations from individual stations.

[0035] For example, the preset time period could be from August 1, 2023 to August 31, 2024. The time granularity could be, but is not limited to, 1 hour, 3 hours, or 1 day. As an example, the technical roadmap for a radiation forecast correction method based on multi-source data fusion and feature distillation is as follows: Figure 2 As shown. Figure 2 The diagram illustrates a technical flowchart of some embodiments of the radiation forecast correction method based on multi-source data fusion and characteristic distillation according to this application, showing the entire process from data preparation, residual Kriging correction to characteristic distillation training and real-time inference.

[0036] It should be noted that the aforementioned wireless connection methods may include, but are not limited to, 3G / 4G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wideband) connection, and other currently known or future wireless connection methods.

[0037] Step 102: Based on the site observation dataset, perform physical correction processing on each reanalysis data in the reanalysis dataset to generate privileged information and obtain a privileged information set.

[0038] In some embodiments, the aforementioned execution entity may perform physical correction processing on each reanalysis data in the aforementioned reanalysis dataset based on the aforementioned site observation dataset to generate privileged information and obtain a privileged information set.

[0039] In practice, the aforementioned implementing entity can perform physical correction processing on each reanalysis data in the aforementioned reanalysis dataset based on the aforementioned site observation dataset to generate privileged information through the following steps:

[0040] The first step involves using the station observation data corresponding to the reanalysis data in the aforementioned station observation dataset to perform physical bias correction on the reanalysis data using the residual kriging algorithm, thereby generating a reanalysis correction field that combines high spatial resolution and physical consistency. As an example, the flowchart of the residual kriging algorithm is shown below. Figure 3 As shown. Figure 3 A schematic diagram of the residual kriging correction principle is shown in some embodiments of the radiation forecast correction method based on multi-source data fusion and characteristic distillation according to this application, illustrating the generation process from discrete site residuals to a continuous correction field.

[0041] The second step is to mark the corrected reanalysis correction field as privileged information for the training phase.

[0042] In practice, the aforementioned implementing entity can perform physical bias correction on the reanalysis data using the residual kriging algorithm based on the station observation data corresponding to the reanalysis data in the aforementioned station observation dataset, thereby generating a reanalysis correction field with both high spatial resolution and physical consistency:

[0043] The first step is to determine the residual set between the above-mentioned station observation data and the above-mentioned reanalysis data using the following formula:

[0044] ,

[0045] in, Indicates the serial number. Indicates the first One site, Represents the corresponding sites in the residual set The residual, Indicates the corresponding station in the station observation data The observed values, This indicates that the data was reanalyzed and interpolated to the site. The value of . Here, processing the reanalysis data can be done by processing the radiation included in the reanalysis data. For example, This could involve reanalyzing the data and interpolating it to the location of the station using a bilinear method. ,

[0046] in, , Indicates the interpolation parameters. , , , Indicates site latitude and longitude , This indicates the latitude and longitude of the bottom left corner of the grid cell. , This indicates the latitude and longitude of the upper right corner of the grid cell. , , , This represents the radiation values ​​at four grid points in the reanalysis data.

[0047] The second step is to determine the Euclidean distance set and the radiative semivariance set based on the above residual set using the following formula:

[0048] ,

[0049] in, Indicates the serial number. Indicates the first One site, Indicates the corresponding sites for Euclidean distance. to station The Euclidean distance between them and Indicates site latitude and longitude and Indicates site latitude and longitude Represents the corresponding sites in the radiation semivariogram set. to station The radiation semivariance between them Represents the corresponding sites in the residual set The residual.

[0050] The third step is to determine the average semivariance set based on the above Euclidean distance set and the above radiation semivariance set using the following formula:

[0051] ,

[0052] in, Indicates the serial number. Represents the Euclidean distance set. A European distance, Represents the interval corresponding to the set of mean semivariance Mean semivariance Representing an interval The total number of point pairs within, This represents a constant. Here, the interval... The length is the step size. For example, It can be 1. An interval can be an interval among several intervals that divide the Euclidean distance between all pairs of stations.

[0053] The fourth step is to determine the spherical model based on the aforementioned average semivariance set, objective function, and goodness-of-fit formula. The objective function is:

[0054] ,

[0055] in, Describe the objective function. Indicates weight, , Represents the theoretical semivariogram. This indicates the nugget value of the spherical model. Indicates the value of the partial sill. Indicates a change in range.

[0056] The goodness-of-fit formula above is:

[0057] ,

[0058] in, Indicates the goodness of fit. Represents the fitted parameters, , This represents the average of the individual mean semivariances in the mean semivariance set. Represents the value of the semi-mutation function.

[0059] The above spherical model is:

[0060] ,

[0061] in, Represents the value of the semi-mutation function. This represents the target distance value.

[0062] In practice, firstly, the aforementioned executing entity can set initial parameters. , , Secondly, the aforementioned executing entity can construct the objective function using the average semivariogram set and the theoretical semivariogram. Then, the executing entity can estimate the initial parameters using the weighted least squares method. , , Then, in response to determining that the goodness of fit is less than the preset goodness value, the aforementioned executing entity adjusts the initial parameters. , , Finally, in response to determining that the goodness of fit is greater than or equal to a preset goodness value, the aforementioned executing entity sets the adjusted initial parameters as parameters to determine the spherical model. For example, the preset goodness value could be 0.7. Here, the spherical model is a semi-variogram model commonly used in Kriging interpolation. The spherical model can be a pair of points formed within the interval. The model to be fitted.

[0063] The fifth step is to determine the grid residual field based on the aforementioned spherical model. In practice, firstly, a set of grid distances corresponding to the reanalysis data is determined using a preset distance algorithm. For example, the preset distance algorithm could be the Euclidean distance algorithm. The grid distances in the grid distance set can represent the distances from grid points to stations. Therefore, the distance between each unobserved grid point and each known station can be calculated. Secondly, for each grid distance in the aforementioned grid distance set, the grid distance is input into the aforementioned spherical model to obtain the grid residual values. Then, the obtained grid residual values ​​are determined as the grid residual field.

[0064] Step 6: Based on the above gridded residual field, the reanalysis correction field, which combines high spatial resolution and physical consistency, is determined using the following formula:

[0065] ,

[0066] in, This represents a reanalysis correction field that combines high spatial resolution with physical consistency. This indicates the background field of the reanalyzed data. This represents the gridded residual field. Here, the background field of the reanalysis data can be the reanalysis data itself.

[0067] Therefore, the bias field can be superimposed on the original reanalysis field, and spatial upsampling can be performed using a bicubic interpolation algorithm to align its resolution with historical NWP data, thus obtaining the corrected reanalysis field.

[0068] Therefore, before model fusion, the grid-scale bias of the coarse-resolution reanalysis data can be corrected using site observations, providing a more reliable background field for the model, reducing errors at the source, and using the corrected reanalysis field as privileged information.

[0069] Step 103: Train a multi-source fusion correction model based on the historical NWP dataset and privileged information set.

[0070] In some embodiments, the aforementioned execution entity can train a multi-source fusion correction model based on the aforementioned historical NWP dataset and the aforementioned privileged information set. The multi-source fusion correction model can be a pre-trained model that takes NWP data as input and outputs a high-precision radiometric correction prediction field. As an example, the model structure of the multi-source fusion correction model can refer to... Figure 4 . Figure 4 The diagram shows the model structure of a multi-source fusion correction model according to some embodiments of the radiation forecast correction method based on multi-source data fusion and feature distillation according to this application, detailing the Teacher-Student branch, the embedding location of SEBlock (channel attention mechanism module), and the scope of the loss function.

[0071] In practice, the aforementioned implementing entity can train a multi-source fusion correction model based on the aforementioned historical NWP dataset and privileged information set through the following steps:

[0072] The first step involves partitioning the aforementioned historical NWP dataset and privilege information set to generate a training sample set. The training samples in the training sample set include sample NWP data and sample privilege information. In practice, the first sub-step involves, for each historical NWP data point in the historical NWP dataset, firstly, the executing entity can identify the historical NWP data as sample NWP data. Secondly, the executing entity can identify the privilege information corresponding to the historical NWP data in the privilege information set as sample privilege information. Then, the executing entity can identify the sample NWP data and sample privilege information as training samples. The second sub-step involves the executing entity defining each identified training sample as the training sample set.

[0073] The second step is to determine the initial multi-source fusion correction model. This initial model can include an initial teacher branch neural network model and an initial student branch neural network model. Both models are based on an encoder-decoder architecture. The initial teacher branch neural network model takes sample NWP data and sample privileged information (7 input channels, including radiation, temperature, air pressure, specific humidity, total cloud cover, low cloud cover, and sample privileged information) as input and outputs the initial teacher correction field. The initial student branch neural network model takes sample NWP data (6 input channels, including radiation, temperature, air pressure, specific humidity, total cloud cover, and low cloud cover) as input and outputs the initial high-precision radiation correction forecast field. In practice, the output of the initial teacher branch neural network model (the initial teacher correction field) can be distilled into the initial student branch neural network model. Therefore, by minimizing the difference between the intermediate layer feature maps of the teacher and student branches during the training phase, the latent features in the historical reanalysis data can be distilled into the student branch, allowing the student branch to output correction results based solely on NWP data during the inference phase.

[0074] The initial teacher branch neural network model may include: an initial teacher encoder, an initial teacher channel attention module, and an initial teacher decoder.

[0075] The initial teacher encoder can consist of four convolutional modules, each containing a "Conv3x3-BN-ReLU (3*3 convolution-batch normalization-corrected linear unit)" layer and a max-pooling layer. Initial teacher channel attention modules (SE Blocks) are embedded at the bottlenecks of layers 2, 3, and 4. For example, the SE Blocks compress the spatial dimension through global average pooling and then generate channel weights through two fully connected layers (FC). A reduction ratio of r=16 is set to enable the model to automatically identify the feature channels with the greatest impact on radiation (such as cloud water content) and suppress noise channels.

[0076] The initial teacher channel attention module can include a global average pooling layer and two fully connected layers. For example, its reduction ratio r is set to 8-32, preferably 16, a range determined through ablation experiments: when r < 8, the model overfits; when r > 32, channel attention fails; and when r = 16, the minimum mean absolute error (MAE) of radiation forecasts is achieved on the validation set. The initial teacher channel attention module is used to dynamically adjust the feature channel weights according to the physical importance of meteorological factors to enhance the response values ​​of cloud microphysical feature channels that are sensitive to radiation attenuation.

[0077] The initial teacher decoder can be a U-Net network.

[0078] Furthermore, a modal dropout mechanism is introduced into the initial teacher branch neural network model. This modal dropout mechanism can be as follows: First, a preset dropout probability is set. Second, in each training batch, a random mask variable M∈{0,1} following a Bernoulli distribution is generated, where p(M=0)=p, and p represents the preset dropout probability. This random mask variable M is applied to the sample privileged information channel of the initial teacher branch neural network model. When M=1, the input of the initial teacher branch neural network model includes sample NWP data and sample privileged information. When M=0, the channel value corresponding to the sample privileged information in the input tensor is forcibly set to zero, retaining only the sample NWP data channel, thus simulating a data-deficient environment during the inference stage. As an example, setting the preset dropout probability to 0.5, the input of the initial teacher branch neural network model has a 50% probability of being complete data containing both sample NWP data and sample privileged information (at this time, the model learns to use dual-source information for high-precision reconstruction), and a 50% probability of being sample NWP data (at this time, the model is forced to use only sample NWP data and its contextual information to recover the feature distribution provided by the sample privileged information).

[0079] The initial student branch neural network model may include: an initial student encoder, an initial student channel attention module, an initial convolutional layer, and an initial student decoder.

[0080] The initial student encoder can consist of four sets of convolutional modules, each containing a "Conv3x3-BN-ReLU (3*3 convolution-batch normalization-corrected linear unit)" and a max pooling layer. The initial student channel attention module (SE Block) is embedded at the bottleneck of layers 2, 3, and 4.

[0081] The initial student channel attention module may include a global average pooling layer and two fully connected layers, with a reduction ratio r set to 8-32, preferably 16. This range is determined through ablation experiments: when r < 8, the model overfits; when r > 32, the channel attention fails; and when r = 16, the minimum mean absolute error (MAE) of the radiation forecast is obtained on the validation set.

[0082] The initial convolutional layer can be a 1*1 convolutional layer, used to map the feature dimensions of the feature map output by the initial student channel attention module to be consistent with the features of the feature map output by the initial teacher channel attention module, so as to calculate the feature distillation loss.

[0083] The initial student decoder can be a U-Net network.

[0084] Therefore, the initial teacher channel attention module and the initial student channel attention module are configured to adaptively weight the radiation-sensitive meteorological element channels such as cloud cover and water vapor based on the global statistical characteristics of the input NWP meteorological variables (including cloud cover, specific humidity, temperature, air pressure, and radiation). This is used to dynamically adjust the feature channel weights according to the physical importance of meteorological factors, so as to enhance the response value of the cloud microphysical feature channels that are sensitive to radiation attenuation.

[0085] The third step is to select target training samples from the aforementioned training sample set. In practice, the executing entity can randomly select training samples from the training sample set as target training samples.

[0086] The fourth step involves inputting the sample privilege information and NWP data from the selected target training samples into the initial multi-source fusion correction model to obtain the initial high-precision radiometric correction prediction field. In practice, the aforementioned execution entity can input the sample privilege information and NWP data from the selected target training samples into the initial multi-source fusion correction model through the following sub-steps to obtain the initial high-precision radiometric correction prediction field:

[0087] The first sub-step involves inputting the sample privilege information and sample NWP data from the selected target training samples into the initial teacher branch neural network model included in the initial multi-source fusion correction model, thereby obtaining the initial teacher correction field. Specifically, when inputting the sample privilege information and sample NWP data into the initial teacher branch neural network model, the input value of the sample privilege information is set to zero with a preset dropout probability to simulate a data-deficient environment during the inference stage. For example, the preset dropout probability could be 0.5.

[0088] The second sub-step involves inputting the selected target training samples, including the sample NWP data, into the initial student branch neural network model included in the initial multi-source fusion correction model, and distilling the initial teacher correction field into the initial student branch neural network model to obtain the initial high-precision radiation correction prediction field.

[0089] The fifth step involves determining the difference between the initial high-precision radiometric correction prediction field and the sample privilege information included in the selected target training samples, based on a preset loss function. This difference can include: sparse site loss, spatial gradient difference loss, feature distillation loss, and structural similarity loss. The preset loss function can be:

[0090] ,

[0091] in, Indicates the difference value. Indicates sparse site loss, This represents the spatial gradient difference loss. Indicates characteristic distillation loss, Represents structural similarity loss. , , , The weighted hyperparameters representing the various loss terms, for example, It can be 1.0. It can be 0.5. It can be 0.1. It could be 0.2;

[0092] The sparse site loss is generated using the following formula:

[0093] ,

[0094] in, Indicates the serial number. Indicates the number of radiation observation stations. Indicates the corresponding initial high-precision radiometric correction prediction field The predicted value at that location, The corresponding information in the sample privilege information The true value of the measured radiation at the location;

[0095] The spatial gradient difference loss is determined by calculating the gradient difference between the predicted field (initial high-precision radiometrically corrected forecast field) and the ground truth field (sample privileged information) in the longitude and latitude directions. The gradient calculation uses the Sobel operator and is generated by the following formula:

[0096] ,

[0097] in, This represents the total number of pixels in the radiative grid field (sample privileged information). Indicates the prediction field. Represents the truth field. This represents the Sobel gradient operator along the longitude direction. This represents the Sobel gradient operator along the latitudinal direction. Represents the L1 norm;

[0098] The characteristic distillation loss is generated using the following formula:

[0099] ,

[0100] in, The feature map represents the initial student branch neural network model (the feature map of the output of the initial convolutional layer included in the initial student branch neural network model). The feature map represents the initial teacher branch neural network model (the feature map output by the initial teacher channel attention module included in the initial teacher branch neural network model). Represents the L2 norm;

[0101] The structural similarity loss is generated using the following formula:

[0102] ,

[0103] in, Representation of structural similarity algorithms, This represents the corresponding image patch extracted from the prediction field (initial high-precision radiometrically corrected prediction field). This represents the corresponding image patch extracted from the truth field (sample privilege information). This represents the mean of the predicted field image patch. The mean of the image patch representing the truth field. This represents the standard deviation of the predicted field image patch. This represents the standard deviation of the true field image patch. This represents the covariance between the image patches representing the predicted field and the true field. and Here, represents a constant. , This represents the maximum radiance value (the largest radiance value among all radiance values ​​included in the sample privilege information). For example, the corresponding image patch extracted from the prediction field could be a 9x9 image patch extracted from the prediction field. Similarly, the corresponding image patch extracted from the ground truth field could be a 9x9 image patch extracted from the ground truth field.

[0104] Therefore, firstly, sparse site loss can be used to ensure the accuracy of site locations. Secondly, spatial gradient difference loss can be used to maintain the consistency between the spatial gradient of the output field and the input forecast, thus preserving details. Next, structural similarity loss can be used to ensure that the output field is consistent with the large-scale structure of the reanalysis data. Then, feature distillation loss can be used to constrain the output of the auxiliary decoder to be as similar as possible to the output of the main fusion network (U-Net), so as to distill the knowledge of the "teacher network" (main U-Net) that integrates multi-source information into the "student path" (backbone encoder + auxiliary decoder) that relies solely on forecast data.

[0105] Step 6: In response to the determination that the difference value meets the preset adjustment conditions, adjust the network parameters of the initial multi-source fusion correction model. The preset adjustment conditions may be: there are no consecutive 20 epochs (the model completes one forward and backward propagation process on the entire training dataset) where the difference value does not decrease, and the number of epochs is less than 200. In practice, in response to the determination that the difference value meets the preset adjustment conditions, the above-mentioned execution entity can adjust the network parameters of the initial multi-source fusion correction model using the Adam (Adaptive Moment Estimation) algorithm. The initial learning rate of the Adam algorithm can be 0.001. The Adam algorithm uses an exponential decay strategy for learning rate scheduling, that is, the learning rate is multiplied by 0.5 every 50 consecutive epochs where the validation loss does not decrease.

[0106] Optionally, the aforementioned execution entity may also determine the initial multi-source fusion correction model as the trained multi-source fusion correction model in response to the determination that the difference value does not meet the aforementioned preset adjustment conditions.

[0107] Therefore, after the multi-source fusion correction model is trained, only the new high-resolution numerical weather prediction product (single channel) needs to be input into the real-time operational forecast. The input data flows through the backbone encoder and auxiliary decoder, which already possess "distillation knowledge" (or directly uses the main U-Net encoder and a lightweight decoder), and can directly output the corrected high-precision gridded field without any reanalysis data or real-time dense site observations.

[0108] Step 104: Preprocess the acquired real-time NWP data to generate real-time NWP preprocessed data.

[0109] In some embodiments, the execution entity can preprocess the acquired real-time NWP data to generate real-time NWP preprocessed data. The real-time NWP data can be NWP data at the current time acquired from a terminal device via a wired or wireless connection. In practice, the execution entity can preprocess the acquired real-time NWP data using a preset preprocessing algorithm to generate real-time NWP preprocessed data. For example, the preset preprocessing algorithm can be a spatiotemporal interpolation algorithm.

[0110] Step 105: Input the real-time NWP preprocessed data into the pre-trained multi-source fusion correction model to obtain a high-precision radiation correction prediction field.

[0111] In some embodiments, the aforementioned execution entity can input the real-time NWP preprocessed data into a pre-trained multi-source fusion correction model to obtain a high-precision radiometric correction prediction field. As an example, a comparison diagram of the data flow during the training and inference phases of the multi-source fusion correction model is shown below. Figure 5 As shown, Figure 5The diagram shows a comparison of the data flow during the training and inference phases of a multi-source fusion correction model according to some embodiments of the radiation forecast correction method based on multi-source data fusion and feature distillation of this application, highlighting the zero dependence of the inference phase on privileged information.

[0112] Therefore, firstly, for sparsely supervised image-to-image conversion, a U-Net network is used to process the entire meteorological field, and a composite loss function is designed to effectively utilize the ground truth of sparse stations while ensuring the spatial continuity and physical rationality of the output across the entire field. Secondly, a feature distillation mechanism is introduced: by discarding inputs and using an auxiliary decoder, the core problem of inconsistent input data during training and deployment is explicitly solved, ensuring the model's practicality and robustness in operational environments. Next, through multi-scale feature fusion and attention mechanisms, the natural structure of U-Net is used to capture features from meteorological to local scales, and the importance of features is dynamically adjusted through an attention mechanism to improve correction accuracy.

[0113] As an example, this application uses historical data from a land region (70°E (longitude) - 140°E, 10°N (latitude) - 50°N), covering 13 months from August 1, 2023 to August 31, 2024, for experimental purposes. The data is divided chronologically: August 2023 - May 2024 serves as the training set, June - July 2024 as the validation set, and August 2024 as the test set, for training the multi-source fusion correction model. Experimental results show that, on the independent test set (August 2024), the method of this invention outperforms the original forecast, the traditional model output statistics (MOS) method, and ordinary convolutional neural networks (CNN) on multiple metrics. The evaluation results are as follows: Figure 6 As shown. Figure 6 The diagram shows a comparison of the original forecast, conventional MOS, ordinary CNN, a radiation forecast correction method based on multi-source data fusion and feature distillation, and the evaluation results assuming that ERA5 can acquire them in real time. Figure 6 In this context, MAE represents the mean absolute error, RMSE represents the root mean square error, and W / m 2 This refers to tiles per square meter.

[0114] The various embodiments of this application have the following beneficial effects: Firstly, the radiation forecast correction method based on multi-source data fusion and feature distillation, as described in some embodiments of this application, effectively fuses multi-source information and learns under the supervision of station ground truth. This reduces the mean absolute error (MAE) and root mean square error (RMSE) of the corrected radiation forecast at the validation stations, while increasing the correlation coefficient, thus significantly improving forecast accuracy. Secondly, the innovative feature distillation mechanism ensures the reliability of the model in real-world operational scenarios, overcoming the bottleneck of inconsistent training and deployment data, and enabling the automatic generation of high-precision correction products for operational use. Furthermore, the multi-source fusion correction model outputs a spatially continuous and physically reasonable gridded field, overcoming the shortcomings of traditional point-to-point correction methods in field structure, and exhibiting reasonable performance even in areas without stations. Next, the gradient loss and other design features enable the model to retain valuable small-scale weather details from the original high-resolution forecast while correcting biases. Therefore, a high-precision, high-quality spatial field with good detail preservation can be obtained for the corrected fine-resolution field.

[0115] Further reference Figure 7 As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of a radiation forecast correction device based on multi-source data fusion and characteristic distillation. These embodiments of the radiation forecast correction device based on multi-source data fusion and characteristic distillation are similar to... Figure 1 Corresponding to the method embodiments shown, this radiation forecast correction device based on multi-source data fusion and feature distillation can be specifically applied to various electronic devices.

[0116] like Figure 7 As shown, a radiation forecast correction device 700 based on multi-source data fusion and feature distillation in some embodiments includes: an acquisition unit 701, a physical correction unit 702, a training unit 703, a preprocessing unit 704, and an input unit 705. The acquisition unit 701 is configured to acquire a reanalysis dataset, a site observation dataset, and a historical NWP dataset; the physical correction unit 702 is configured to perform physical correction processing on each reanalysis data in the reanalysis dataset based on the site observation dataset to generate privileged information, thus obtaining a privileged information set; the training unit 703 is configured to train a multi-source fusion correction model based on the historical NWP dataset and the privileged information set; the preprocessing unit 704 is configured to preprocess the acquired real-time NWP data to generate real-time NWP preprocessed data; and the input unit 705 is configured to input the real-time NWP preprocessed data into the pre-trained multi-source fusion correction model to obtain a high-precision radiation correction forecast field.

[0117] It is understandable that the units and references described in the radiation forecast correction device 700 based on multi-source data fusion and characteristic distillation are... Figure 1The steps in the described method correspond accordingly. Therefore, the operations, features, and beneficial effects described above for the method are also applicable to the radiation forecast correction device 700 based on multi-source data fusion and feature distillation and the units contained therein, and will not be repeated here.

[0118] This application also provides a computer device 800. For example... Figure 8 As shown, the computer device 800 includes a bus 801, a processor 802, a memory 803, and a communication interface 804. The processor 802, the memory 803, and the communication interface 804 communicate with each other via the bus 801. The computer device 800 can be a server or a terminal device. It should be understood that this application does not limit the number of processors and memories in the computer device 800.

[0119] The 801 bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 8 The bus 801 may be represented by a single line, but this does not mean that there is only one bus or one type of bus. The bus 801 may include a path for transmitting information between various components of the computer device 800 (e.g., memory 803, processor 802, communication interface 804).

[0120] Processor 802 may include any one or more processors such as a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor (MP), or a digital signal processor (DSP).

[0121] Memory 803 may include volatile memory, such as random access memory (RAM). Memory 803 may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid state drive (SSD).

[0122] The memory 803 stores executable program code, and the processor 802 executes this executable program code to implement the functions of the aforementioned acquisition unit, physical correction unit, training unit, preprocessing unit, and input unit, thereby realizing the radiation forecast correction method based on multi-source data fusion and feature distillation. That is, the memory 803 stores instructions for executing the aforementioned radiation forecast correction method based on multi-source data fusion and feature distillation.

[0123] The communication interface 804 uses transceiver modules such as, but not limited to, network interface cards and transceivers to enable communication between the computer device 800 and other devices or communication networks.

[0124] This application also provides a chip, which includes a processor and a data interface. The processor reads instructions stored in the memory through the data interface to execute the above-described radiation forecast correction method based on multi-source data fusion and feature distillation.

[0125] This application also provides a computer-readable storage medium. The computer-readable storage medium can be any available medium capable of being stored by a computing device, or a data storage device such as a data center containing one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive). The computer-readable storage medium includes instructions that instruct the computing device to execute the aforementioned radiation forecast correction method based on multi-source data fusion and feature distillation.

[0126] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0127] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of this application.

Claims

1. A radiation forecast correction method based on multi-source data fusion and characteristic distillation, characterized in that, include: Obtain the reanalysis dataset, site observation dataset, and historical NWP dataset; Based on the site observation dataset, physical correction is performed on each reanalysis data in the reanalysis dataset to generate privileged information, thus obtaining a privileged information set; Based on the historical NWP dataset and the privileged information set, a multi-source fusion correction model is trained. The acquired real-time NWP data is preprocessed to generate real-time NWP preprocessed data; The real-time NWP preprocessed data is input into a pre-trained multi-source fusion correction model to obtain a high-precision radiation correction prediction field; The step of training a multi-source fusion correction model based on the historical NWP dataset and the privileged information set includes: The historical NWP dataset and the privilege information set are partitioned to generate a training sample set, wherein the training samples in the training sample set include: sample NWP data and sample privilege information; The initial multi-source fusion correction model is determined, which includes: an initial teacher branch neural network model and an initial student branch neural network model; Select target training samples from the training sample set; The selected target training samples, including sample privilege information and sample NWP data, are input into the initial multi-source fusion correction model to obtain the initial high-precision radiometric correction prediction field. When the sample privilege information and sample NWP data are input into the initial teacher branch neural network model, the input value of the sample privilege information is set to zero with a preset discard probability to simulate the data missing environment in the inference stage. Based on a preset loss function, the difference between the initial high-precision radiometric correction prediction field and the sample privilege information included in the selected target training samples is determined. The difference values ​​include: sparse site loss, spatial gradient difference loss, feature distillation loss, and structural similarity loss. In response to the determination that the difference value meets the preset adjustment conditions, the network parameters of the initial multi-source fusion correction model are adjusted.

2. The radiation forecast correction method based on multi-source data fusion and characteristic distillation according to claim 1, characterized in that, The step of performing physical correction processing on each reanalysis data in the reanalysis dataset based on the station observation dataset to generate privileged information includes: Based on the station observation data in the station observation dataset that corresponds to the reanalysis data, the reanalysis data is subjected to physical bias correction processing using the residual kriging algorithm to generate a reanalysis correction field with both high spatial resolution and physical consistency. The corrected reanalysis correction field is marked as privileged information during the training phase.

3. The radiation forecast correction method based on multi-source data fusion and characteristic distillation according to claim 2, characterized in that, The reanalysis data, based on the station observation data corresponding to the reanalysis data in the station observation dataset, undergoes physical bias correction processing using the residual kriging algorithm to generate a reanalysis correction field with both high spatial resolution and physical consistency, including: The residual set between the site observation data and the reanalysis data is determined using the following formula: ， in, Indicates the serial number. Indicates the first One site, Indicates the corresponding sites in the residual set The residual, Indicates the corresponding station in the station observation data The observed values, This indicates that the data was reanalyzed and interpolated to the site. The value; Based on the residual set, the Euclidean distance set and the radiative semivariance set are determined using the following formula: ， in, Indicates the serial number. Indicates the first One site, Indicates the corresponding sites for Euclidean distance. Arrive at the station The Euclidean distance between them and Indicates site latitude and longitude and Indicates site latitude and longitude Represents the corresponding sites in the radiation semivariogram set. Arrive at the station The radiation semivariance between them Indicates the corresponding sites in the residual set The residual; Based on the Euclidean distance set and the radiation semivariance set, the average semivariance set is determined using the following formula: ， in, Indicates the serial number. Represents the Euclidean distance set. A European distance, Represents the interval corresponding to the set of mean semivariance Mean semivariance Representing an interval The total number of point pairs within, Represents a constant; Based on the aforementioned average semivariance set, objective function, and goodness-of-fit formula, a spherical model is determined, wherein the objective function is: ， in, Describe the objective function. Indicates weight, , Represents the theoretical semivariogram. This indicates the nugget value of the spherical model. Indicates the value of the partial sill. Indicates a change in range. The goodness-of-fit formula is: ， in, Indicates the goodness of fit. Represents the fitted parameters, , This represents the average of the individual mean semivariances in the mean semivariance set. Represents the value of the semi-mutation function. The spherical model is: ， in, Represents the value of the semi-mutation function. Indicates the target distance value; Based on the spherical model, determine the grid residual field; Based on the gridded residual field, the reanalysis correction field, which combines high spatial resolution and physical consistency, is determined by the following formula: ， in, This represents a reanalysis correction field that combines high spatial resolution with physical consistency. This indicates the background field of the reanalyzed data. This represents the lattice residual field.

4. The radiation forecast correction method based on multi-source data fusion and characteristic distillation according to claim 1, characterized in that, The method further includes: In response to the determination that the difference value does not meet the preset adjustment condition, the initial multi-source fusion correction model is determined as the trained multi-source fusion correction model.

5. The radiation forecast correction method based on multi-source data fusion and characteristic distillation according to claim 1, characterized in that, The preset loss function is: ， in, Indicates the difference value. Indicates sparse site loss, This represents the spatial gradient difference loss. Indicates characteristic distillation loss, Represents structural similarity loss. , , , This represents the weighted hyperparameters for each loss term; The sparse site loss is generated using the following formula: ， in, Indicates the serial number. Indicates the number of radiation observation stations. Indicates the corresponding initial high-precision radiometric correction prediction field The predicted value at that location, The corresponding information in the sample privilege information The true value of the measured radiation at the location; The spatial gradient difference loss is determined by calculating the gradient difference between the predicted field and the true field in the longitude and latitude directions. The gradient calculation uses the Sobel operator and is generated by the following formula: ， in, This represents the total number of pixels in the radiative grid field. Indicates the prediction field. Represents the truth field, This represents the Sobel gradient operator along the longitude direction. This represents the Sobel gradient operator along the latitudinal direction. Represents the L1 norm; The characteristic distillation loss is generated using the following formula: ， in, The feature map representing the initial student branch neural network model. The feature map representing the initial teacher branch neural network model. Represents the L2 norm; The structural similarity loss is generated using the following formula: ， in, Representation of structural similarity algorithms, This represents the corresponding image patch extracted from the prediction field. This represents the corresponding image patch extracted from the truth field. This represents the mean of the predicted field image patch. The mean of the image patch representing the truth field. This represents the standard deviation of the predicted field image patch. This represents the standard deviation of the true field image patch. This represents the covariance between the image patches representing the predicted field and the true field. and Here, represents a constant. , This indicates the maximum radiation value.

6. A radiation forecast correction device based on multi-source data fusion and characteristic distillation, characterized in that, include: The acquisition unit is configured to acquire reanalysis datasets, site observation datasets, and historical NWP datasets; The physical correction unit is configured to perform physical correction processing on each reanalysis data in the reanalysis dataset based on the site observation dataset to generate privileged information and obtain a privileged information set; The training unit is configured to train a multi-source fusion correction model based on the historical NWP dataset and the privileged information set; The training unit is further configured as follows: The historical NWP dataset and the privilege information set are partitioned to generate a training sample set, wherein the training samples in the training sample set include: sample NWP data and sample privilege information; The initial multi-source fusion correction model is determined, which includes: an initial teacher branch neural network model and an initial student branch neural network model; Select target training samples from the training sample set; The selected target training samples, including sample privilege information and sample NWP data, are input into the initial multi-source fusion correction model to obtain the initial high-precision radiometric correction prediction field. When the sample privilege information and sample NWP data are input into the initial teacher branch neural network model, the input value of the sample privilege information is set to zero with a preset discard probability to simulate the data missing environment in the inference stage. Based on a preset loss function, the difference between the initial high-precision radiometric correction prediction field and the sample privilege information included in the selected target training samples is determined. The difference values ​​include: sparse site loss, spatial gradient difference loss, feature distillation loss, and structural similarity loss. In response to the determination that the difference value meets the preset adjustment conditions, the network parameters of the initial multi-source fusion correction model are adjusted. The preprocessing unit is configured to preprocess the acquired real-time NWP data to generate real-time NWP preprocessed data. The input unit is configured to input the real-time NWP preprocessed data into a pre-trained multi-source fusion correction model to obtain a high-precision radiometric correction prediction field.

7. A computer device, characterized in that, The computer device includes a processor, a memory, and a computer program stored in the memory and executable by the processor, wherein when the computer program is executed by the processor, it implements the steps of the method as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-5.

Images